Stays within observable conversation scope
Does the response stay within information that's observable from the conversation? Does it introduce unverifiable policies, features, or specifics not discussed?
This metric works with RAG-based AI systems where knowledge bases may be incomplete. Making ungrounded claims (even if true elsewhere) creates confusion when users can't verify the information.
